Fundamentals

Architecture

Neutron AI sits beneath your application, agents, reasoning providers, simulators, and optimisation solvers as a persistent intelligence layer. Your platform keeps ownership of users, authorisation, provider credentials, billing, tool policy, final model calls, and consequential actions. Neutron receives only the scoped memory, knowledge, decision, and outcome data you choose to send.

flowchart TB
  A["Your app or agent"] --> B["Neutron REST, SDK, CLI, or MCP"]
  B --> C["Context Engine"]
  C --> D["Memory engine"]
  C --> E["Temporal knowledge graph"]
  C --> F["Decision records"]
  D --> G["Consequence and simulation"]
  E --> G
  F --> G
  G --> H["Observed outcomes and reviewed learning"]
  C --> I["Your model or reasoning provider"]
  F --> J["Your verified solver"]

Responsibilities

Layer	Owned by	Responsibility
Product identity	Your platform	Users, workspaces, roles, billing, and authorization decisions.
Neutron boundaries	Shared contract	Nucleus and Scope IDs that keep memory, graph knowledge, decisions, outcomes, simulations, and learning separated.
Memory and knowledge	Neutron AI	Remember, recall, version entities and relationships, preserve provenance, manage contradictions, and tombstone deleted records.
Context Engine	Neutron AI	Assemble eligible memory, graph state, decisions, consequences, outcomes, procedures, contradictions, and reviewed learning within a token budget.
Decision intelligence	Shared contract	Your application supplies objectives, constraints, options, evidence, and actors; Neutron preserves trade-offs, snapshots, consequences, and outcomes.
Provider requests	Your platform	Model selection, provider credentials, tool policy, final prompts, and response handling.
Solvers and simulators	Your platform or specialist provider	Calculate domain-specific feasible or optimal results and calibrated scenarios within an explicit proof or validation boundary.
MCP access	Your platform and Neutron AI	You decide which tools a host can call and which scopes the token can access.

Read Path

sequenceDiagram
  participant App as Your app
  participant N as Neutron AI
  participant M as Model provider

  App->>N: Request context for a task, entities, and timeframe
  N-->>App: Return bounded memory, graph, decisions, outcomes, warnings, and trace
  App->>M: Send provider request with current user input and context
  M-->>App: Return model response
  App->>N: Optionally record the decision and observed outcome

Write Path

Use remember when your product has a durable fact, preference, rule, procedure, decision, outcome, or summary worth reusing. Use entity and relationship writes when stable identities, attributes, graph dimensions, or historical validity matter. Record predicted consequences separately from observations, and activate learning only after the required evidence review. Use tombstone operations when a user revokes, deletes, or corrects knowledge.

Integration Boundary

Neutron AI does not need your provider API keys, raw session cookies, payment details, or private application secrets. Keep those inside your own server boundary and pass Neutron only scoped, minimised content necessary for future agent work. Neutron is not the model, solver, professional decision-maker, or external-action authority.

Fundamentals

Memory model